Lab play estimates an upper bound on model performance through 24 target entities, each more difficult than the previous one.

“That gives you an easy way how to evaluate the upper bound on the performance because we have 24 target entities and each new entity is more difficult than the previous one.”